Instruction-Level Parallelism 1

Report
Instruction-Level Parallelism
compiler techniques and branch prediction
prepared and Instructed by
Shmuel Wimer
Eng. Faculty, Bar-Ilan University
March 2014
Instruction-Level Parallelism 1
1
Concepts and Challenges
The potential overlap among instructions is called
instruction-level parallelism (ILP).
Two approaches exploiting ILP:
•
•
Hardware discovers and exploit the parallelism dynamically.
Software finds parallelism, statically at compile time.
CPI for a pipelined processor:
Ideal pipeline CPI + Structural stalls + Data hazard stalls +
Control stalls
Basic block: a straight-line code with no branches.
• Between three to six instructions typical size .
• Too small to exploit significant amount of parallelism.
• We must exploit ILP across multiple basic blocks.
March 2014
Instruction-Level Parallelism 1
2
March 2014
Instruction-Level Parallelism 1
3
Loop-level parallelism exploits parallelism among
iterations of a loop. A completely parallel loop adding
two 1000-element arrays:
Within an iteration there is no opportunity for overlap,
but every iteration can overlap with any other iteration.
The loop can be unrolled either statically by compiler or
dynamically by hardware.
Vector processing is also possible. Supported in DSP,
graphics, and multimedia applications.
March 2014
Instruction-Level Parallelism 1
4
Data Dependences and Hazards
If two instructions are parallel, they can be executed
simultaneously in a pipeline without causing any
stalls, assuming the pipeline has sufficient resources.
Two dependent instructions must be executed in
order, but can often be partially overlapped.
Three types of dependences: data dependences,
name dependences, and control dependences.
Instruction j is data dependent on instruction i if:
• i produces a result that may be used by j, or
• j is data dependent on an instruction k, and k is data
dependent on i.
March 2014
Instruction-Level Parallelism 1
5
The following loop increments a vector of values in
memory starting at 0(R1), with the last element at
8(R2)), by a scalar in register F2.
The data dependences in this code sequence involve
both floating-point and integer data.
Since between two data dependent instructions there
is a chain of one or more data hazards, they cannot
execute simultaneously or be completely overlapped.
March 2014
Instruction-Level Parallelism 1
6
Data dependence conveys:
•
•
•
the possibility of a hazard,
the order in which results must be calculated, and
an upper bound on how much parallelism can be exploited.
Detecting dependence registers is straightforward.
• Register names are fixed in the instructions.
Dependences that flow through memory locations are
more difficult to detect.
• Two addresses may refer to the same location but look
different: For example, 100(R4) and 20(R6).
• The effective address of a load or store may change from one
execution of the instruction to another (so that 20(R4) and
20(R4) may be different).
March 2014
Instruction-Level Parallelism 1
7
Name Dependences
A name dependence occurs when two instructions use
the same register or memory location, called a name,
but there is no flow of data between the instructions.
If i precedes j in program order:
Anti dependence between i and j occurs when j
writes a register or memory location that i reads.
The original ordering must be preserved to ensure
that i reads the correct value.
March 2014
Instruction-Level Parallelism 1
8
Output dependence occurs when i and j write the
same register or memory location. Their ordering must
be preserved to ensure proper value written by j .
Name dependence is not a true dependence.
• The instructions involved can execute simultaneously or be
reordered.
• The name (register # or memory location) is changed so the
instructions do not conflict.
Register renaming can be more easily done.
• Done either statically by a compiler or dynamically by the
hardware.
March 2014
Instruction-Level Parallelism 1
9
Data Hazards
A hazard is created whenever a dependence between
instructions is close enough.
• Program order must be preserved.
The goal of both SW and HW techniques is to exploit
parallelism by preserving program order only where it
affects the outcome of the program.
• Detecting and avoiding hazards ensures that necessary
program order is preserved.
Data hazards are classified depending on the order of
read and write accesses in the instructions. Consider
two instructions i and j, with i preceding j
March 2014
Instruction-Level Parallelism 1
10
The possible data hazards are:
RAW (read after write). j tries to read a source before i
writes it.
• The most common, corresponding to a true data dependence.
• Program order must be preserved.
WAW (write after write). j tries to write an operand
before it is written by i.
•
•
•
Writes are performed in the wrong order, leaving the value
written by i rather than by j.
Corresponds to an output dependence.
Present only in pipelines that write in more than one pipe
stage or allow an instruction to proceed even when a
previous instruction is stalled.
March 2014
Instruction-Level Parallelism 1
11
WAR (write after read). j tries to write a destination
before it is read by i, so i incorrectly gets the new
value.
•
•
•
•
Arises from an anti dependence.
Cannot occur in most static issue pipelines because all
reads are early (in ID) and all writes are late (in WB).
Occurs when there are some instructions that write results
early in the pipeline and other instructions that read a
source late in the pipeline.
Occurs also when instructions are reordered.
March 2014
Instruction-Level Parallelism 1
12
Control Dependences
A control dependence determines the ordering of i
with respect to a branch so that the i is executed in
correct order and only when it should be.
There are two constraints imposed by control
dependences:
• An instruction that is control dependent on a branch cannot
be moved before the branch so that its execution is no longer
controlled by the branch.
• An instruction that is not control dependent on a branch
cannot be moved after the branch so that its execution is
controlled by the branch.
March 2014
Instruction-Level Parallelism 1
13
Consider this code:
If we do not maintain the data dependence involving
R2, the result of the program can be changed.
If we ignore the control dependence and move the
load before the branch, the load may cause a memory
protection exception. (why?)
No data dependence prevents interchanging the
BEQZ and the LW; it is only the control dependence.
March 2014
Instruction-Level Parallelism 1
14
Compiler Techniques for Exposing ILP
Pipeline is kept full by finding sequences of unrelated
instructions that can be overlapped in the pipeline.
To avoid stall, a dependent instruction must be
separated from the source by a distance in clock
cycles equal to the pipeline latency of that source.
Example: Latencies of FP operations
March 2014
Instruction-Level Parallelism 1
15
Code adding
scalar to vector:
Straightforward MIPS assembly code:
R1 is initially the address top element in the array.
F2 contains the scalar value s.
R2 is pre computed, so that 8(R2) is the array bottom.
March 2014
Instruction-Level Parallelism 1
16
Without any scheduling
the loop takes 9 cycles:
Scheduling the loop
obtains only two stalls,
taking 7 cycles:
March 2014
Instruction-Level Parallelism 1
17
The actual work on the array is just 3/7 cycles (load,
add, and store). The other 4 are loop overhead. Their
elimination requires more operations relative to the
overhead.
Loop unrolling replicating the loop body multiple times
can be used.
• Adjustment of the loop termination code is required.
• Used also to improve scheduling.
Instruction replication alone with usage of same
registers could prevent effective scheduling. Different
registers for each replication are required (increasing
the required number of registers).
March 2014
Instruction-Level Parallelism 1
18
Unrolled code (not rescheduled)
1 stall
2 stalls
1 stall
2 stalls
1 stall
2 stalls
1 stall
2 stalls
1 stall
Stalls are still there.
Run in 27 clock cycles, 6.75 per element.
March 2014
Instruction-Level Parallelism 1
19
Unrolled and rescheduled code
No stalls required.
Execution dropped 14
clock cycles, 3.5 per
element.
Compared with 9 per
element before
unrolling or scheduling
and 7 when scheduled
but not unrolled.
March 2014
Instruction-Level Parallelism 1
20
Problem: The number of loop iterations n is usually
unknown. We would like to unroll the loop to make k
copies of its body.
Two consecutive loops are generated Instead.
The first executes n mod k times and has a body that
is the original loop.
The second is the unrolled body surrounded by an
outer loop that iterates n/k times.
For large n, most of the execution time will be spent
in the unrolled loop body.
March 2014
Instruction-Level Parallelism 1
21
Branch Prediction
performance losses can be reduced by predicting how
branches will behave.
Branch prediction (BP) Can be done statically at
compilation and dynamically at execution time.
The simplest static scheme is to predict a branch as
taken. Misprediction equal to the untaken frequency
(34% for the SPEC benchmark).
BP based on profiling is more accurate.
March 2014
Instruction-Level Parallelism 1
22
Misprediction on SPEC92 for a profile-based predictor
March 2014
Instruction-Level Parallelism 1
23
Dynamic Branch Prediction
The simplest is a BP buffer, a small 1-bit memory
indexed by the lower portion of the address of the
branch instruction.
Useful only to reduce the branch delay when it is longer
than the time to compute the possible target PCs.
BP may have been put there by another branch that
has the same low-order address bits!
Fetching begins in the predicted direction. If it was
wrong, the BP bit is inverted and stored back.
March 2014
Instruction-Level Parallelism 1
24
Problem: Even if almost always taken, We will likely
predict incorrectly twice. (Why?)
Solution: 2-bit BP
are often used. It
must miss twice
before it is changed.
An n-bit counter is possible. If counter > 2n-1 – 1, taken
is predicted; otherwise, untaken.
2-bit do almost as well, thus used by most systems.
March 2014
Instruction-Level Parallelism 1
25
Correlating Branch Predictors
2-bit uses only the recent behavior of a single branch
for BP. Accuracy can be improved if the recent
behavior of other branches are considered.
Consider the code:
Let aa and bb be assigned to registers R1 and R2, and
label the three branches b1, b2, and b3. The compiler
generates the typical MIPS code:
March 2014
Instruction-Level Parallelism 1
26
The behavior of b3 is correlated with that of b1 and b2.
A predictor using only the behavior of a single branch
to predict its outcome is blind of this behavior.
Correlating or two-level predictors add information
about the most recent branches to decide on a branch.
March 2014
Instruction-Level Parallelism 1
27
An (m,n) BP uses the last m branches to choose from
2m n-bit BPs for a single branch.
More accurate than 2-bit and requires simple HW.
The global history of the most recent m branches is
recorded in an m-bit shift register. The buffer is indexed
using a concatenation of the low-order branch address
bits with the m-bit history.
The 6-bit index of a 64 entries (2,2) buffer is formed by
the 4 low-order bits of the branch address + 2 global
bits obtained from the two most recent branches.
March 2014
Instruction-Level Parallelism 1
28
For a fair comparison of the performance of BPs, the
same number of state bits are used.
The number of bits in an (m,n) predictor is:
2m × n × # of prediction entries.
A 2-bit predictor w/o global history is a (0,2) predictor.
Example: How many bits are in the (0,2) BP with 4K
entries? How many entries are in a (2,2) predictor with
the same number of bits?
A 4K-entries BP has 20 × 2 × 4K = 8K bits.
A (2,2) BP having a total of 8K bits satisfies
22 × 2 × # of prediction entries = 8K.
The # of prediction entries is therefore = 1K.
March 2014
Instruction-Level Parallelism 1
29
not much
improvement
significant
improvement
March 2014
Instruction-Level Parallelism 1
30
Tournament Predictors
Tournament predictors combine predictors based on
global and local information.
They achieve better accuracy and effectively use very
large numbers of prediction bits.
Tournament BPs use a 2-bit saturating counter per
branch to select between two different BP (local,
global), based on which was most effective in recent
predictions.
As in a simple 2-bit predictor, the saturating counter
requires two mispredictions before changing the
identity of the preferred BP.
March 2014
Instruction-Level Parallelism 1
31
March 2014
Instruction-Level Parallelism 1
32

similar documents